サッカー選手のアノテーションの品質向上を目指した深層学習のためのデータセットの構築およびパラメータの最適化に関する研究

抄録

In recent years, object detection methods using deep learning have been utilized for player position analysis in sports information science, contributing to performance improvement. However, these methods require a lot of effort in creating training datasets. The annotation system developed by the authors previously improved the efficiency of datasets generation, but full automation was not achieved. Therefore, the present research proposes a semi-automatic generation method for training datasets using image processing technology and AI. And then, we implemented a method to refine the training dataset while building the unique detection model, taking into account the background difference and detection results from the AI model. Ultimately, it was confirmed that as the unique model is refined, it will also be able to semi-automatically generate datasets for learning, contributing to the improvement of the annotation system.